Multi-modal joint spectrum allocation method and system

Abstract

The application provides a multi-modal joint spectrum allocation method and system, relating to the technical field of data processing, comprising: acquiring multi-modal observation data in low-altitude airspace, respectively encoding the multi-modal observation data according to modes to obtain communication mode potential features and radar mode potential features; performing intermediate fusion on the communication mode potential features and the radar mode potential features at a feature layer, and imposing information constraints on the intermediate fusion result to obtain fusion features; measuring the distribution difference between the communication mode potential features and the radar mode potential features, and performing distribution uniformization processing on the fusion features based on the obtained distribution difference measurement result to obtain distribution-uniformized fusion features; constructing a low-altitude spectrum resource three-dimensional situation map based on the distribution-uniformized fusion features; generating a spectrum resource dynamic allocation strategy based on the low-altitude spectrum resource three-dimensional situation map, and issuing the spectrum resource dynamic allocation strategy to an execution object.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a multimodal joint spectrum allocation method and system. Background Technology

[0002] With the rapid development of the low-altitude economy, application scenarios such as drone logistics, low-altitude inspection, urban air traffic, emergency rescue, low-altitude security, and agricultural and forestry plant protection continue to expand. The deployment density of various wireless communication and sensing devices in the low-altitude airspace is rapidly increasing, and the low-altitude electromagnetic environment is characterized by multi-source heterogeneity, dynamic changes, and high complexity.

[0003] In low-altitude airspace, low-altitude aircraft typically rely on various wireless signals, such as cellular mobile communication, WiFi, dedicated short-range communication, and telemetry and remote control links, for command and control and service transmission. Meanwhile, the detection and identification of low-altitude targets requires the use of active sensing devices such as radar to generate echo observation information. Therefore, the spectrum usage status in low-altitude airspace is no longer something that can be independently characterized by a single frequency band or a single device, but rather manifests as a comprehensive spectrum situational issue spanning multiple frequency bands, protocols, and platforms.

[0004] However, most existing low-altitude spectrum management and wireless resource scheduling methods are still based on a single communication standard, fixed frequency band monitoring, or simple energy detection to carry out spectrum sensing. When faced with a large amount of heterogeneous observation data in a complex electromagnetic environment, the accuracy and stability of the spectrum sensing results are difficult to meet the requirements, and the sensing results usually remain at the state recognition level, which is difficult to directly support the spectrum resource scheduling needs of low-altitude flight missions.

[0005] Therefore, how to improve the accuracy and stability of spectrum sensing in complex low-altitude electromagnetic environments and realize the dynamic scheduling of spectrum resources is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] This invention provides a multimodal joint spectrum allocation method and system to solve the problem of how to improve the accuracy and stability of spectrum sensing in complex low-altitude electromagnetic environments and realize dynamic scheduling of spectrum resources in the prior art.

[0007] This invention provides a multimodal joint spectrum allocation method, comprising: Multimodal observation data in the low-altitude airspace is acquired, and the multimodal observation data is feature-encoded according to mode to obtain communication mode potential features and radar mode potential features; wherein, the multimodal observation data includes multi-band communication signals and radar echo information; The communication mode latent features and the radar mode latent features are fused at the feature layer, and information constraints are applied to the fusion result to obtain fused features; wherein, the information constraints are used to enhance the correlation between the fused features and the target task output, and to reduce the redundant components in the fused features. The distribution difference between the communication mode potential features and the radar mode potential features is measured, and the fused features are subjected to distribution unification processing based on the obtained distribution difference measurement results to obtain the distribution unified fused features; Based on the fusion features after the distribution is unified, a three-dimensional situation map of low-altitude spectrum resources is constructed with spatial location, time and frequency as coordinate axes. Based on the three-dimensional situation map of low-altitude spectrum resources, a dynamic spectrum resource allocation strategy is generated and the dynamic spectrum resource allocation strategy is sent to the execution objects.

[0008] According to a multimodal joint spectrum allocation method provided by the present invention, the distribution difference between the latent features of the communication mode and the latent features of the radar mode is measured, and the fused features are subjected to distribution unification processing based on the obtained distribution difference measurement results, including: The statistical distribution difference between the latent features of the communication mode and the latent features of the radar mode in a unified feature space is measured by at least one of the optimal transmission distance, kernel method distance and adversarial distribution matching method, and the distribution difference measurement result is obtained. With minimizing the distribution difference metric as the optimization objective, the representations of the communication mode latent features and the radar mode latent features in the unified feature space are aligned.

[0009] According to a multimodal joint spectrum allocation method provided by the present invention, the distribution difference between the latent features of the communication modes and the latent features of the radar modes is measured, including: The latent features of the communication modes and the latent features of the radar modes are respectively subjected to frequency domain transformation to obtain the transform domain features corresponding to each mode; The dominant frequency domain components in the transform domain features are retained, while the non-dominant frequency domain components are suppressed to obtain the frequency domain compressed features. A random mapping matrix with restricted isometry is used to perform dimensionality reduction projection on the frequency domain compressed features to obtain dimensionality reduction features located in a low-dimensional space. The distribution difference is approximated based on the dimensionality reduction features in the low-dimensional space to obtain the distribution difference measurement result.

[0010] According to a multimodal joint spectrum allocation method provided by the present invention, the intermediate fusion of the communication mode latent features and the radar mode latent features at the feature layer includes: Based on the fusion weights corresponding to each mode, the latent features of the communication mode and the latent features of the radar mode are fused using at least one of the following methods: weighted fusion, gated fusion, attention fusion, and tensor fusion. The fusion weight is either a preset static weight or a dynamic weight adaptively generated based on dynamic weight generation factors. The dynamic weight generation factors include at least one of the following: current link quality, modal confidence, noise level, electromagnetic congestion in the flight area, and radar target tracking stability.

[0011] According to the multimodal joint spectrum allocation method provided by the present invention, the target task output includes at least one of spectrum occupancy identification results, interference detection results, available channel prediction results, low-altitude communication link quality assessment results, and flight area spectrum conflict early warning results. The reduction of redundant components in the fused features includes: suppressing duplicate information and local anomalous observation components between the communication mode latent features and the radar mode latent features.

[0012] According to the multimodal joint spectrum allocation method provided by the present invention, after the step of distributing the spectrum resource dynamic allocation strategy to the execution object, the method further includes: The execution status feedback data generated by the execution object in response to the dynamic spectrum resource allocation strategy, as well as newly added multimodal observation data, are obtained; wherein, the execution status feedback data includes at least one of frequency band switching results, link quality changes, and interference avoidance results; The execution status feedback data and the newly added multimodal observation data are used together as the input data for the next cycle. The feature encoding, intermediate fusion, distribution unification processing, and construction of the three-dimensional situation map of low-altitude spectrum resources are performed in sequence to obtain the updated three-dimensional situation map of low-altitude spectrum resources. Based on the updated three-dimensional situation map of low-altitude spectrum resources, an updated dynamic spectrum resource allocation strategy is generated and distributed to the execution object.

[0013] According to a multimodal joint spectrum allocation method provided by the present invention, the step of constructing a three-dimensional situation map of low-altitude spectrum resources based on the fused features after distribution unification, using spatial location, time, and frequency as coordinate axes, includes: The fused features after the distribution is made consistent are mapped onto a low-altitude spatial grid with spatial location, time and frequency as coordinate axes; Situational attribute information is superimposed on the low-altitude spatial grid to construct a three-dimensional situational map of the low-altitude spectrum resources; wherein, the situational attribute information includes at least one of spectrum occupancy intensity, interference level, link quality index, target activity density, and resource conflict degree.

[0014] According to the multimodal joint spectrum allocation method provided by the present invention, the dynamic spectrum resource allocation strategy includes at least one of the following: priority channel recommendation, frequency band switching, link switching, transmit power adjustment, beam control, access mode adjustment, interference avoidance path suggestion, and mission area resource reconfiguration. The target of execution includes at least one of low-altitude aircraft, ground base stations, and edge control platforms.

[0015] According to a multimodal joint spectrum allocation method provided by the present invention, the step of performing feature encoding on the multimodal observation data according to each mode includes: For the multi-band communication signal, encoding is performed based on at least one of time-domain waveform, frequency-domain spectrum, time-frequency joint representation and modulation structure information to obtain the potential features of the communication mode. The radar echo information is encoded based on at least one of range dimension information, velocity dimension information, angle dimension information, micro-motion dimension information and trajectory sequence information to obtain the radar modal potential features; The feature encoding network used includes at least one or a combination of convolutional networks, temporal networks, graph structure modeling networks, and transform domain feature mapping networks.

[0016] According to a multimodal joint spectrum allocation method provided by the present invention, before performing feature encoding on the multimodal observation data according to each mode, the method further includes: The multimodal observation data is time-aligned and spatially registered to obtain spatiotemporally aligned multimodal observation data; The step of performing feature encoding on the multimodal observation data according to each mode includes: performing feature encoding on the spatiotemporally aligned multimodal observation data according to each mode to obtain the communication mode potential features and the radar mode potential features with a unified dimension.

[0017] According to the multimodal joint spectrum allocation method provided by the present invention, the multi-band communication signal includes at least one of: cellular mobile communication signal, WiFi signal, dedicated short-range communication signal, and UAV remote control and telemetry link signal; The radar echo information includes at least one of the following: target echo intensity, range, Doppler frequency shift, angular spectrum, micro-motion characteristics, and point trajectory characteristics.

[0018] The present invention also provides a multimodal joint spectrum allocation system, comprising: A modal coding module is used to acquire multimodal observation data in the low-altitude airspace, and to perform feature coding on the multimodal observation data according to mode to obtain communication mode potential features and radar mode potential features; wherein, the multimodal observation data includes multi-band communication signals and radar echo information; The intermediate fusion and constraint module is used to perform intermediate fusion of the communication mode potential features and the radar mode potential features at the feature layer, and to apply information constraints to the intermediate fusion result to obtain fused features; wherein, the information constraints are used to enhance the correlation between the fused features and the target task output, and to reduce the redundant components in the fused features. The distribution unification module is used to measure the distribution difference between the communication mode potential features and the radar mode potential features, and to perform distribution unification processing on the fused features based on the obtained distribution difference measurement results to obtain the distribution unified fused features. The situation map construction module is used to construct a three-dimensional situation map of low-altitude spectrum resources based on the fused features after the distribution is consistent, with spatial location, time and frequency as coordinate axes; The dynamic allocation module is used to generate a dynamic spectrum resource allocation strategy based on the three-dimensional situation map of low-altitude spectrum resources, and to distribute the dynamic spectrum resource allocation strategy to the execution object.

[0019] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the multimodal joint spectrum allocation method as described above.

[0020] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the multimodal joint spectrum allocation method as described above.

[0021] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the multimodal joint spectrum allocation method as described above.

[0022] The multimodal joint spectrum allocation method and system provided by this invention acquires multimodal observation data, including multi-band communication signals and radar echo information, in the low-altitude airspace. After feature encoding by mode, intermediate fusion is performed at the feature layer, establishing a coupling relationship between communication and sensing information in the latent space. Compared to merging results at the decision stage, this method retains more fine-grained complementary information related to spectrum occupancy, interference conditions, and link quality. By imposing information constraints on the intermediate fusion results, the fused features more centrally express the discriminative information related to the target task and suppress redundant components, improving the accuracy and robustness of sensing results in complex electromagnetic environments. Furthermore, by integrating the latent features of communication modes with radar echo information... The distribution differences among the potential features of different modes are measured and the fused features are processed to achieve distribution consistency. This alleviates the representation offset problem caused by differences in physical mechanisms and statistical laws of heterogeneous modes, and enhances the model's generalization ability across devices, regions, and environments. Furthermore, a three-dimensional situation map of low-altitude spectrum resources is constructed using spatial location, time, and frequency as coordinate axes. Based on the three-dimensional situation map of low-altitude spectrum resources, a dynamic spectrum resource allocation strategy is generated and distributed to the execution object. This enables the perception results to be directly converted into executable scheduling actions, forming a continuous processing link from perception to decision-making to execution. This improves the utilization efficiency of low-altitude airspace wireless resources, communication reliability, and spectrum scheduling response capability. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating the multimodal joint spectrum allocation method provided by the present invention; Figure 2 This is a schematic diagram of intermediate fusion and information constraint provided by the present invention; Figure 3 This is a schematic diagram of the three-dimensional situation map of low-altitude spectrum resources provided by the present invention; Figure 4 This is a schematic diagram of the dynamic spectrum allocation closed loop provided by the present invention; Figure 5 This is a schematic diagram of distribution uniformity and low-complexity distance estimation provided by the present invention; Figure 6 This is a block diagram of the overall structure of the multimodal joint spectrum allocation system for low-altitude economy provided by the present invention; Figure 7This is a flowchart illustrating the multimodal joint spectrum allocation method for low-altitude economy provided by the present invention. Figure 8 This is a schematic diagram of the multimodal joint spectrum allocation system provided by the present invention; Figure 9 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0026] The execution entities of the multimodal joint spectrum allocation method for low-altitude economy provided in this invention include, but are not limited to, the following: spectrum management controllers deployed at ground base stations, control towers, or low-altitude monitoring stations; edge control platforms or edge computing servers deployed at edge nodes; embedded processors deployed on airborne platforms; central servers or cluster servers of UAV dispatch centers; and industrial control computers, cloud server clusters, or personal computers with computing capabilities. In the following embodiments, for ease of description, the edge control platform is consistently used as the execution entity. Those skilled in the art will understand that the steps in the following embodiments can also be implemented by other types of execution entities.

[0027] The following description uses an urban low-altitude drone logistics scenario as a general example: In a core area A of a city, several drone delivery stations, multiple ground-based 5G / 6G base stations, several WiFi access points, dedicated short-range communication gateways, and millimeter-wave radars deployed at control towers and edge nodes are deployed. Multiple drones simultaneously perform parcel delivery, low-altitude inspection, and emergency delivery tasks within area A. The edge control platform needs to intelligently sense and dynamically allocate low-altitude spectrum resources in this scenario.

[0028] Figure 1 This is a flowchart illustrating the multimodal joint spectrum allocation method provided by the present invention, as shown below. Figure 1 As shown, the method includes the following: Step S101: Acquire multi-modal observation data in the low-altitude airspace, and perform feature encoding on the multi-modal observation data according to mode to obtain communication mode potential features and radar mode potential features; wherein, the multi-modal observation data includes multi-band communication signals and radar echo information; The edge control platform acquires multimodal observation data in the low-altitude airspace, which includes multi-band communication signals and radar echo information.

[0029] Multimodal observation data refers to the collection of electromagnetic observation information originating from different physical mechanisms and observation modes. Among them, multi-band communication signals refer to wireless communication electromagnetic signals covering different frequency bands received by radio frequency front-ends deployed at ground base stations, towers, edge nodes, low-altitude monitoring stations, or airborne platforms. These signals can characterize spectrum occupancy, link activity, and interference distribution characteristics in different airspaces and time periods. Radar echo information refers to target reflected echo signals or derived data based on these echoes, transmitted by radar transmitters deployed at the aforementioned nodes and received by their receivers. These signals can reflect target position, speed, trajectory changes, and echo structure characteristics.

[0030] The acquisition methods may include, but are not limited to: pulling observation data in real time from the distributed radio frequency front-end and radar front-end through wired backhaul network interface, wireless backhaul link, dedicated low-altitude communication channel or edge buffer interface; or reading the collected and pre-stored observation samples from storage devices.

[0031] In the pervasive embodiment, the edge control platform synchronously collects communication signals from 5G / 6G base stations, WiFi access points, and dedicated short-range communication gateways within region A via fiber optic backhaul links, while simultaneously acquiring echo signals from tower millimeter-wave radar and edge node radar via dedicated interfaces.

[0032] The edge control platform performs feature encoding on the acquired multimodal observation data according to the mode to obtain the potential features of the communication mode and the potential features of the radar mode.

[0033] Feature encoding refers to the process of mapping raw observation data from its original physical space, such as time-domain sampling sequences, frequency-domain spectra, distance-Doppler maps, etc., into a vectorized representation with a unified dimension and a unified semantic scale.

[0034] Communication mode latent features refer to the latent vectors formed after multi-band communication signals are encoded and processed, which can be jointly analyzed with radar modes; radar mode latent features refer to the latent vectors formed after radar echo information is encoded and processed, which have the same or compatible dimensions as communication mode latent features.

[0035] Modal-specific feature encoding can be achieved by: constructing separate encoding network branches for communication signals and radar echoes, with each branch receiving input for its own modality and outputting a latent vector of fixed dimensions; or by having encoding networks sharing some parameters work together. Each encoding branch can employ a network structure capable of feature extraction.

[0036] In the through-through embodiment, the communication signal is input into the communication coding branch after obtaining the time-spectrum diagram via short-time Fourier transform, and the radar echo is input into the radar coding branch after range-Doppler processing; both coding branches output... d dimensional latent vector, d The preset potential spatial dimension.

[0037] Assume the system receives a total of M The modal input, the first m Each modality is processed by its corresponding encoder to obtain a latent feature representation. ,in .

[0038] Step S102: At the feature layer, the communication mode potential features and the radar mode potential features are fused in the intermediate layer, and information constraints are applied to the intermediate fusion result to obtain the fused features; Information constraints are used to enhance the correlation between fused features and the output of the target task, and to reduce redundant components in the fused features.

[0039] Intermediate fusion refers to a processing method that jointly models the features of each modality in the latent feature layer after encoding. Figure 2 This is a schematic diagram of the intermediate fusion and information constraint provided by the present invention, as shown below. Figure 2 As shown, multimodal data input is mapped to a unified latent space by an independent modal encoder to obtain communication latent features and radar latent features; the two enter the intermediate fusion module for feature layer weighting or attention fusion to obtain fused features z; fused features z are linearly or nearly linearly mapped by the target mapping layer, which is simultaneously constrained by the information constraint module to maximize task relevance, and finally outputs the enhanced fused feature representation.

[0040] Fusion features refer to the joint representation formed by the intermediate fusion of communication mode potential features and radar mode potential features, which is used to support downstream sensing tasks and resource decisions.

[0041] In one implementation, the fusion feature can be represented as: in, For the first m The fusion weights of each modality satisfy... .

[0042] Information constraints refer to loss terms or regularization terms applied to the mapping path from fused features to the target task output, used to optimize the efficiency of feature representation.

[0043] The constraint serves two purposes: firstly, it enhances the correlation between the fused features and the outputs of the target task, such as spectrum occupancy, interference risk, and link quality; secondly, it suppresses intermodal redundancy and noise components. The constraint can be applied by introducing a correlation loss term between the fused features and the target task label, jointly optimizing the task loss and constraint loss during the training phase. The target task output refers to the decision result of the downstream sensing task performed based on the fused features.

[0044] Step S103: Measure the distribution difference between the communication mode potential features and the radar mode potential features, and perform distribution unification processing on the fused features based on the obtained distribution difference measurement results to obtain the distribution unification fused features.

[0045] Distribution difference refers to the deviation between the statistical distribution of communication mode potential features and radar mode potential features in a unified feature space.

[0046] Because communication signals and radar echoes differ significantly in physical mechanisms, sampling frequencies, noise models, and spatial coverage characteristics, they often exhibit different distribution patterns in a unified feature space after being encoded.

[0047] Distribution difference measurement results refer to scalar or vector results obtained by quantifying the above distribution deviations through statistical / geometric measurement methods.

[0048] Distribution uniformity processing refers to the process of optimizing the coding network parameters or latent features themselves with the goal of minimizing the distribution difference metric, so that the distribution of communication mode latent features and radar mode latent features tends to be consistent in a unified feature space. After distribution uniformity processing, the fused features are the distribution-uniformed fused features.

[0049] Step S104: Based on the fusion features after distribution unification, construct a three-dimensional situation map of low-altitude spectrum resources with spatial location, time and frequency as coordinate axes.

[0050] A three-dimensional situation map of low-altitude spectrum resources refers to a visualized or data-driven structure formed by mapping the results of multimodal fusion sensing onto the joint spatiotemporal-frequency resource representation in the low-altitude airspace, using the three dimensions of spatial location, time, and frequency as the basic coordinate axes.

[0051] Figure 3 This is a schematic diagram of the three-dimensional situation map of low-altitude spectrum resources provided by the present invention, as shown below. Figure 3 As shown, the construction of a three-dimensional situation map of low-altitude spectrum resources uses spatial location, time, and frequency as basic input coordinate axes, and overlays information such as spectrum occupancy, interference distribution, and resource availability to output a three-dimensional situation map of low-altitude spectrum resources.

[0052] The construction method may include mapping the fused features after distribution uniformity to pre-divided low-altitude spatial grids, time slices and frequency band grids, with the value of each grid point reflecting the resource usage status under the location-time-frequency band combination.

[0053] In the comprehensive embodiment, the edge control platform spatially divides region A into a 100m×100m×50m cubic grid, divides time slices into 10-second intervals, and divides frequency bands into 20MHz granularities, thereby constructing a description of the spectrum resource usage status on any spatial grid, any time slice, and any frequency band grid point within region A.

[0054] Step S105: Based on the three-dimensional situation map of low-altitude spectrum resources, generate a dynamic spectrum resource allocation strategy and distribute the dynamic spectrum resource allocation strategy to the execution object.

[0055] Dynamic spectrum resource allocation strategy refers to a set of executable scheduling instructions generated for aircraft, ground base stations, or edge nodes in low-altitude airspace, based on the resource usage status reflected in the three-dimensional situation map of low-altitude spectrum resources. These instructions include spectrum usage, link adjustment, power management, beam control, and regional coordination. The execution object refers to the physical node that receives and executes the above-mentioned dynamic spectrum resource allocation strategy.

[0056] Figure 4 This is a schematic diagram of the dynamic spectrum allocation closed loop provided by the present invention, as shown below. Figure 4 As shown, the multimodal spectrum sensing results are constructed into a three-dimensional spectrum situation map and then enter the spectrum resource strategy generation stage. The output includes strategies such as channel selection, frequency band switching, transmit power adjustment, access mode switching, interference avoidance, and resource reconfiguration. These strategies are then issued to execution objects, including low-altitude aircraft, base stations, and edge control platforms, in the form of scheduling instructions.

[0057] The methods of issuing commands may include: issuing commands to the aircraft via a low-altitude communication link, issuing configuration parameters to the ground base station via the base station management interface, or issuing coordination strategies to other edge nodes via the edge control interface. In a general implementation, the edge control platform issues scheduling commands containing suggested operating frequency bands, alternative channels, maximum transmit power, and recommended access base stations to multiple logistics drones in area A.

[0058] In this application, by fusing multi-band communication signals and radar echoes at the feature layer and suppressing redundancy through information constraints, and further reducing cross-modal distribution deviation through distribution consistency, an executable dynamic allocation strategy is generated on a three-dimensional resource representation with spatial location, time, and frequency as coordinate axes. This can improve the accuracy, robustness, and real-time performance of spectrum sensing in complex low-altitude electromagnetic environments and form a complete spectrum management link, supporting intelligent sensing and efficient scheduling of spectrum resources in low-altitude economic scenarios.

[0059] Optionally, the distribution difference between the communication mode latent features and the radar mode latent features is measured, and the fused features are subjected to distribution unification processing based on the obtained distribution difference measurement results, including: The statistical distribution difference between the latent features of the communication mode and the latent features of the radar mode in a unified feature space is measured by at least one of the optimal transmission distance, kernel method distance and adversarial distribution matching method, and the distribution difference measurement result is obtained. With minimizing the distribution difference metric as the optimization objective, the representations of the communication mode latent features and the radar mode latent features in the unified feature space are aligned.

[0060] In this application, optimal transmission distance refers to the distance between two distributions that measures the transport cost based on optimal transmission theory, such as Wasserstein distance and its various approximations; kernel method distance refers to the distance between two distributions that measures the difference based on regenerating kernel Hilbert space, such as maximum mean difference; adversarial distribution matching method refers to a method that distinguishes the distributions to which different modal features belong through a discriminant network and uses the discriminant error as a signal to measure the distribution difference.

[0061] In the distribution uniformity process, the optimization objective is to minimize the distribution difference metric result, so that the representations of communication mode latent features and radar mode latent features tend to align in a unified feature space.

[0062] In one implementation, the overall optimization objective of the system can be expressed as: in, This indicates the loss of spectrum sensing or resource allocation tasks; Indicates information constraint terms; Represents the distribution uniformization term; 、 These are weighting parameters. If the optimal transmission distance is used, then... The design is the sum of the intermodal Wasserstein distances or their approximate forms.

[0063] In the full-scale implementation, the edge control platform uses Wasserstein distance as... In each training batch, a set of samples is sampled from the latent features of the communication mode and the latent features of the radar mode, respectively. The Wasserstein distance between the empirical distributions corresponding to the two sets of samples is estimated and used as the basis for determining the training batch. Substitute it into the overall optimization objective, and and Cooperative minimization is used to update the parameters of each coding branch and fusion module.

[0064] By introducing a distribution distance metric with good mathematical properties and using minimizing this metric as the optimization objective for mode alignment, the distribution offset between communication modes and radar modes in a unified feature space can be effectively mitigated, thereby improving the stability of the fused representation and its generalization ability to unknown low-altitude scenarios.

[0065] Optionally, the distribution difference between the communication mode latent features and the radar mode latent features is measured, including: The latent features of the communication modes and the latent features of the radar modes are respectively subjected to frequency domain transformation to obtain the transform domain features corresponding to each mode; The dominant frequency domain components in the transform domain features are retained, while the non-dominant frequency domain components are suppressed to obtain the frequency domain compressed features. A random mapping matrix with restricted isometry is used to perform dimensionality reduction projection on the frequency domain compressed features to obtain dimensionality reduction features located in a low-dimensional space. The distribution difference is approximated based on the dimensionality reduction features in the low-dimensional space to obtain the distribution difference measurement result.

[0066] Figure 5 This is a schematic diagram of the distribution uniformity and low-complexity distance estimation provided by the present invention, as shown below. Figure 5 As shown, the original distribution 1 corresponding to the high-dimensional features of the communication mode and the original distribution 2 corresponding to the high-dimensional features of the radar mode are processed by frequency domain transformation, preservation of dominant frequency domain components and frequency domain compression, random mapping dimensionality reduction and restricted equidistant dimensionality reduction to obtain feature representations in low-dimensional space. Based on this, low-dimensional distribution distance estimation is performed, and finally the aligned distribution after distribution unification is output.

[0067] The edge control platform performs frequency domain transformations on the latent features of communication modes and radar modes respectively to obtain the transform domain features corresponding to each mode. Frequency domain transformation refers to the reversible transformation that maps latent features from the original sample / time series space to the frequency space, such as discrete Fourier transform, discrete cosine transform, or wavelet transform.

[0068] The process involves preserving the dominant frequency domain components of the transform domain features and suppressing the non-dominant frequency domain components to obtain compressed frequency domain features. The dominant frequency domain components refer to the frequency components in the transform domain features where energy, amplitude, or variance constitutes the majority. These can be determined through threshold filtering, sorting by the top K components based on energy, or sorting by the top-K components based on amplitude. Non-dominant frequency domain components are set to zero or ignored. This step utilizes the sparsity of latent features in the transform domain to reduce redundant dimensionality.

[0069] By employing a random mapping matrix with restricted isometric properties, the frequency-domain compressed features are projected into a lower-dimensional space to obtain the reduced-dimensional features. A random mapping matrix with restricted isometric properties refers to a random matrix that approximately preserves the vector distance structure within the original space; such matrices can be Gaussian random matrices, Bernoulli random matrices, or sub-Gaussian random matrices that satisfy the Restricted Integrity Perspective (RIP).

[0070] In a low-dimensional space, the distribution difference is approximated based on dimensionality-reduced features to obtain a measure of distribution difference. In a low-dimensional space, the computational cost of calculating the distribution distance can be significantly reduced because the dimensionality is greatly reduced while the features still approximately maintain their original geometric structure.

[0071] In the comprehensive embodiment, the communication mode latent features and radar mode latent features, originally with a dimension of d=512, are first subjected to discrete cosine transform and the frequency components with the highest energy percentage (top 80%) are retained to obtain approximately 160-dimensional frequency domain compressed features. These are then projected onto a 64-dimensional low-dimensional space using a Gaussian random matrix that satisfies the restricted isometric property. Finally, the distance between the two mode distributions is estimated within this 64-dimensional space, and the resulting estimation is used as... The item participates in the optimization of the overall goal.

[0072] In this application, a two-stage process of frequency domain compression followed by random mapping and dimensionality reduction is employed. While preserving the original feature geometry and discriminative meaning information as much as possible, the calculation of distribution distance is moved from a high-dimensional space to a low-dimensional space, significantly reducing the online computational complexity. This meets the requirements for real-time updating of spectrum situation maps and rapid generation of resource scheduling strategies in low-altitude scenarios.

[0073] Optionally, the intermediate fusion of the communication mode latent features and the radar mode latent features at the feature layer includes: Based on the fusion weights corresponding to each mode, the latent features of the communication mode and the latent features of the radar mode are fused using at least one of the following methods: weighted fusion, gated fusion, attention fusion, and tensor fusion. The fusion weight is either a preset static weight or a dynamic weight adaptively generated based on dynamic weight generation factors. The dynamic weight generation factors include at least one of the following: current link quality, modal confidence, noise level, electromagnetic congestion in the flight area, and radar target tracking stability.

[0074] In this application, specifically, the edge control platform is based on the fusion weights corresponding to each modality. For the potential characteristics of communication modes and radar mode potential characteristics The following methods are used for fusion: weighted fusion: summing the latent features of each modality according to their weights; gated fusion: generating switching signals for each modality through a gating network and selectively transmitting modal features; attention fusion: modeling and weighting the relationships between modalities through self-attention or cross-attention mechanisms; tensor fusion: capturing high-order interactions between modalities through outer product or tensor operations.

[0075] Fusion weights There are two ways to obtain the value: During the system deployment phase, fixed weight values ​​are preset for each modality based on prior experience, and these values ​​remain unchanged during the operation phase.

[0076] The dynamic weights are generated adaptively based on factors such as current link quality, modal confidence, noise level, electromagnetic congestion in the flight area, and radar target tracking stability.

[0077] Specifically, when the link quality is good or the mode confidence is high, the weight of the corresponding mode is increased accordingly; when the noise level is high or the observation reliability decreases, the weight of the corresponding mode is decreased accordingly; when the flight area is in a state of electromagnetic congestion, the weight of the radar mode can be appropriately increased to compensate for the degradation of the communication mode; when the radar target tracking stability is low, the weight of the communication mode can be appropriately increased.

[0078] In the comprehensive implementation, when a sub-region within region A experiences electromagnetic congestion due to temporary activity, the edge control platform detects a decrease in communication mode confidence and adaptively lowers the communication mode weight from 0.5 to 0.3, while increasing the radar mode weight to 0.7. This makes the fused features more dependent on radar mode information, thereby maintaining stable perception performance.

[0079] In this application, by providing multiple optional feature layer fusion methods and a static / dynamic switchable weighting mechanism, the modal contribution can be flexibly adjusted according to the real-time electromagnetic situation in low-altitude scenarios, further improving the adaptability of the fusion representation to complex scenarios and the stability of the sensing results under different link states.

[0080] Optionally, the target task output includes at least one of the following: spectrum occupancy identification result, interference detection result, available channel prediction result, low-altitude communication link quality assessment result, and flight area spectrum conflict early warning result; The reduction of redundant components in the fused features includes: suppressing duplicate information and local anomalous observation components between the communication mode latent features and the radar mode latent features.

[0081] The target mission output includes at least one of the following: spectrum occupancy identification results, which refers to the judgment conclusion on whether each frequency band is occupied in each spatial location and time slot; interference detection results, which refers to the judgment conclusion on whether there are abnormal interference sources and the intensity of interference in each frequency band; available channel prediction results, which refers to the prediction of the set of channels available for use in one or more future time slots; low-altitude communication link quality assessment results, which refers to the assessment conclusion on the quality level of the current low-altitude communication link in different regions and frequency bands; and flight area spectrum conflict early warning results, which refers to the early warning of possible spectrum conflict events in the low-altitude flight mission area.

[0082] Reducing redundant components in fused features specifically includes suppressing the following: duplicate information between communication mode potential features and radar mode potential features, such as highly correlated components of the same target in two modes; and local anomalous observation components, such as anomalous features caused by short-term obstruction, transient strong interference, or sensor anomalies.

[0083] In a comprehensive implementation, for a drone swarm mission within region A, the edge control platform outputs in parallel the set of available channels in the sub-region where the swarm is located within the next 30 seconds, whether there is interference in the current frequency band, and the uplink quality level of the drone swarm, and collaboratively generates scheduling instructions based on the above outputs.

[0084] By clearly defining the output formats of multiple target tasks and the specific suppression targets of redundant components, the fused features can simultaneously support multiple downstream sensing tasks and maintain stable discrimination capabilities in low-altitude scenarios with dense multi-target activities and significant modal redundancy, thereby improving the robustness of the system in complex electromagnetic environments.

[0085] Optionally, in this application, the edge control platform acquires execution status feedback data generated by the dynamic allocation strategy of the execution object's response spectrum resources, as well as newly added multimodal observation data.

[0086] Among them, execution status feedback data refers to the strategy execution results and subsequent status information reported by the execution object to the edge control platform after executing the spectrum resource dynamic allocation strategy, including at least one of the following: frequency band switching results, link quality changes, and interference avoidance results; newly added multi-mode observation data refers to the multi-band communication signals and radar echo information that are continuously collected in the new observation period after the spectrum resource dynamic allocation strategy is issued.

[0087] The edge control platform uses the execution status feedback data and the newly added multimodal observation data as input data for the next cycle, and sequentially performs feature encoding, intermediate fusion, distribution consistency processing, and construction of a three-dimensional situation map of low-altitude spectrum resources to obtain an updated three-dimensional situation map of low-altitude spectrum resources.

[0088] Based on the updated three-dimensional situation map of low-altitude spectrum resources, the edge control platform generates an updated dynamic spectrum resource allocation strategy and distributes the updated dynamic spectrum resource allocation strategy to the execution objects.

[0089] like Figure 4 As shown, the status feedback and closed-loop update process feeds back the dynamic adjustment results of the execution object to the multimodal spectrum sensing end, and then through the construction of a three-dimensional spectrum situation map, generation of spectrum resource strategy, issuance of scheduling instructions and dynamic adjustment of the execution object, a continuously running closed loop is formed.

[0090] In the comprehensive embodiment, at time T1, the edge control platform sends a command to a logistics drone in area A to "switch from frequency band F1 to frequency band F2"; after T1+5 seconds, the drone sends back execution status feedback data that "the uplink quality level of frequency band F2 has been improved to L4, and the interference of the original F1 frequency band has been avoided"; the edge control platform uses this feedback data together with the newly acquired multimodal observation data after T1 as the input for the next cycle, updates the three-dimensional situation map of low-altitude spectrum resources in area A, and generates and issues a new dynamic spectrum resource allocation strategy for the drone and related drones in the surrounding area.

[0091] By incorporating the policy execution feedback of the target and the newly added observation data into the perception-fusion-decision chain of the next cycle, a substantial closed loop is constructed from perception to execution and back from execution to perception. This enables spectrum resource allocation to track the rapid changes in low-altitude target activities and the electromagnetic environment in real time, significantly improving the flexibility, response efficiency and guarantee capability of spectrum resource allocation compared to the traditional static spectrum management method.

[0092] Optionally, the step of constructing a three-dimensional situation map of low-altitude spectrum resources based on the fused features after distribution unification, using spatial location, time, and frequency as coordinate axes, includes: The edge control platform maps the fused features, after distribution uniformity, onto a low-altitude spatial grid with spatial location, time, and frequency as coordinate axes. In this step, spatial grids, time slices, and frequency band grids are pre-divided according to the coverage of the low-altitude airspace, task granularity, and computing power. The fused features are then written into the corresponding grids according to their corresponding spatial-temporal-frequency band labels.

[0093] The edge control platform overlays situational attribute information onto the low-altitude spatial grid to construct a three-dimensional situational map of low-altitude spectrum resources.

[0094] Situational attribute information refers to a set of attributes that characterize the resource usage status and risk level of each grid point, including but not limited to at least one of the following: spectrum occupancy intensity, reflecting the degree to which the spectrum of the grid point is used; interference level, reflecting the level of interference intensity present at the grid point; link quality index, reflecting the reliability of the communication link at the grid point; target activity density, reflecting the density of flight target distribution within the spatial range of the grid point; and resource conflict degree, reflecting the tension among multiple users competing for the same frequency band resources at the grid point.

[0095] like Figure 3 As shown, the three-dimensional situation map of low-altitude spectrum resources combines the above-mentioned situation attribute information with the grid points of the three-dimensional coordinate axis in a layered manner to form a comprehensive expression of the usage status of low-altitude airspace space-frequency resources.

[0096] In the comprehensive embodiment, region A is discretized into several cubic grids, several time slices, and several frequency band grid points; the situational attribute information of each grid point is stored in the form of a multi-channel array, wherein the first channel stores the spectrum occupancy intensity, the second channel stores the interference level, the third channel stores the link quality index, the fourth channel stores the target activity density, and the fifth channel stores the resource conflict degree.

[0097] By overlaying multiple situational attribute information onto a three-dimensional grid, the three-dimensional situational map of low-altitude spectrum resources possesses both fine-grained spatial-temporal-frequency dimensions and can directly reflect resource usage status and risk distribution, providing rich and interpretable decision-making basis for the generation of subsequent dynamic allocation strategies.

[0098] Optionally, the dynamic spectrum resource allocation strategy includes at least one of the following: priority channel recommendation, frequency band switching, link switching, transmit power adjustment, beam control, access method adjustment, interference avoidance path suggestion, and mission area resource reconfiguration. The target of execution includes at least one of low-altitude aircraft, ground base stations, and edge control platforms.

[0099] In this application, the dynamic spectrum resource allocation strategy includes at least one of the following: Priority channel recommendation refers to recommending the channel with the highest priority among multiple candidate channels based on a comprehensive score; Frequency band switching refers to instructing the target to switch from the current frequency band to an alternative frequency band; Link switching refers to instructing the execution object to switch from the current access link to another link; Transmit power adjustment refers to instructing the target to adjust its transmit power according to the recommended power value; Beam control refers to instructing nodes with beamforming capabilities to adjust the beam direction or width. Access method adjustment refers to instructing the execution object to switch between different access standards; Interference avoidance path recommendations refer to flight paths that suggest areas of interference to avoid for low-altitude flight missions; Mission area resource reallocation refers to the overall adjustment of spectrum resource allocation schemes for a specific low-altitude mission area.

[0100] The target of execution includes at least one of the following: low-altitude aircraft, such as drones performing logistics, inspection, and emergency tasks; ground base stations, such as 5G / 6G base stations and WiFi access points; and edge control platforms, such as other edge nodes with collaborative decision-making capabilities.

[0101] In a continuous implementation, the edge control platform simultaneously issues the following strategies during a peak task period: priority channel recommendation and frequency band switching instructions for prioritizing the use of channel C5 and alternative C7. A transmission power adjustment command is issued to U2 to reduce the uplink transmission power limit from 20dBm to 15dBm; a beam control command is issued to base station B1 to adjust the beam direction to an azimuth angle of 120°; and a mission area resource reconfiguration command is issued to edge node E1 to designate sub-region A3 as an interference avoidance zone and release all mission area resources from frequency bands F1 to F3.

[0102] By providing multiple levels of strategy types, including coverage channel selection, frequency band switching, link adjustment, power management, beam control, and mission area reconfiguration, and supporting the distribution to various execution objects such as low-altitude aircraft, ground base stations, and edge control platforms, this method enables comprehensive intelligent scheduling of spectrum resources in various typical low-altitude economic scenarios.

[0103] Optionally, the step of performing feature encoding on the multimodal observation data according to each modality includes: For the multi-band communication signal, encoding is performed based on at least one of time-domain waveform, frequency-domain spectrum, time-frequency joint representation and modulation structure information to obtain the potential features of the communication mode. The radar echo information is encoded based on at least one of range dimension information, velocity dimension information, angle dimension information, micro-motion dimension information and trajectory sequence information to obtain the radar modal potential features; The feature encoding network used includes at least one or a combination of convolutional networks, temporal networks, graph structure modeling networks, and transform domain feature mapping networks.

[0104] In this application, the time-domain waveform refers to the time-domain sampling sequence of the communication signal; the frequency-domain spectrum refers to the frequency-domain power spectrum or amplitude spectrum of the communication signal; the time-frequency joint representation refers to the time-frequency diagram obtained by the communication signal through short-time Fourier transform, wavelet transform, etc.; and the modulation structure information refers to the modulation method, constellation structure, and other information of the communication signal.

[0105] The encoding of radar echo information includes at least one of the following: range dimension information, i.e., echo amplitude / phase of different range cells; velocity dimension information, i.e., Doppler frequency shift; angle dimension information, i.e., angle spectrum; micro-motion dimension information, i.e., micro-Doppler characteristics; trajectory sequence information, i.e., target point sequence.

[0106] The encoding network used in feature encoding includes at least one or a combination of the following: convolutional networks, suitable for extracting local features and spatial structure; temporal networks, such as long short-term memory networks, gated recurrent units, Transformers, etc., suitable for processing sequence features; graph structure modeling networks, such as graph convolutional networks, suitable for modeling the relationship between objects; and transform domain feature mapping networks, such as frequency domain learnable filter structures.

[0107] In the general implementation, a two-dimensional convolutional network based on time-frequency graphs is used as the communication coding branch for 5G / 6G communication signals, outputting a 512-dimensional latent vector; for millimeter-wave radar echoes, a hybrid coding branch combining a range-Doppler graph convolutional network and a time-series network for processing point trajectory sequences is used, also outputting a 512-dimensional latent vector.

[0108] By providing multiple optional encoding inputs and encoding network structures for different physical features of different modalities, feature encoding can be flexibly adapted to various communication standards and radar systems, thereby improving the coverage and expressive ability of latent features to the original modal information.

[0109] Optionally, before performing feature encoding on the multimodal observation data according to each modality, the method further includes: The multimodal observation data is time-aligned and spatially registered to obtain spatiotemporally aligned multimodal observation data; The step of performing feature encoding on the multimodal observation data according to each mode includes: performing feature encoding on the spatiotemporally aligned multimodal observation data according to each mode to obtain the communication mode potential features and the radar mode potential features with a unified dimension.

[0110] In this application, time alignment refers to the unified calibration of the timestamps of each modal observation data so that the observation data from different acquisition devices and different acquisition frequencies are aligned according to a unified time reference.

[0111] Possible methods include synchronizing the clocks of each acquisition device based on GNSS timing; resampling or interpolating each mode sample according to a common time window; and aligning the communication samples with the radar echo period as a reference.

[0112] Spatial registration refers to unifying observation data from different acquisition devices into the same spatial coordinate reference system. Methods include: coordinate transformation based on the position and attitude parameters of the acquisition devices; and spatial grid alignment based on the position, altitude, heading, and mission area information of the low-altitude aircraft.

[0113] After completing temporal alignment and spatial registration, spatiotemporally aligned multimodal observation data is obtained. In subsequent steps, when feature encoding is performed separately for each modality, the encoding input is the spatiotemporally aligned multimodal observation data, and the encoding output of each modality has a unified dimension.

[0114] In the comprehensive embodiment, the 5G / 6G base stations, radar nodes and WiFi access points deployed in Area A are all connected to a unified GNSS timing system; the edge control platform resamples each acquisition stream with a common time window of 10ms, and performs a unified transformation of the spatial labels of each modal observation data based on the urban coordinate system.

[0115] By performing temporal alignment and spatial registration before feature encoding, the differences in time reference and spatial reference system between different acquisition devices are eliminated, ensuring that the input data for subsequent steps such as feature encoding, intermediate fusion, and distribution consistency have consistency and alignability, thereby improving the stability and accuracy of the entire processing chain.

[0116] Optionally, the multi-band communication signal includes at least one of the following: cellular mobile communication signal, WiFi signal, dedicated short-range communication signal, and UAV remote control and telemetry link signal; The radar echo information includes at least one of the following: target echo intensity, range, Doppler frequency shift, angular spectrum, micro-motion characteristics, and point trajectory characteristics.

[0117] Multi-band communication signals include at least one of the following: cellular mobile communication signals, such as 4G LTE, 5G NR and future 6G signals; WiFi signals, such as wireless local area network signals based on the IEEE 802.11 series of protocols; dedicated short-range communication signals, such as dedicated communication standard signals such as DSRC and C-V2X; and drone remote control and telemetry link signals, such as point-to-point remote control signals, image transmission link signals, and telemetry downlink signals.

[0118] Radar echo information includes at least one of the following: target echo intensity, reflecting the degree of reflection of incident radar waves by the target; range, reflecting the radial distance of the target from the radar node; Doppler frequency shift, reflecting the radial velocity of the target; angular spectrum, reflecting the azimuth / pitch angle of the target in the radar field of view; micro-motion characteristics, reflecting the micro-Doppler components generated by the movement of the target's own components, such as rotor modulation characteristics; and point-track characteristics, reflecting the continuous change of the target's position within multiple time slices.

[0119] In the comprehensive embodiment, the multi-band communication signals in region A cover the 5G NR band, the 2.4GHz / 5GHz WiFi band, the 5.9GHz dedicated short-range communication band, and the 900MHz / 2.4GHz UAV remote control and telemetry band; the radar echo information covers the target echo intensity, range, Doppler frequency shift, angle spectrum, rotor micro-motion characteristics, and continuous multi-frame point trajectory characteristics.

[0120] In this application, by specifying the specific types of multi-band communication signals and radar echo information, the method can cover a wide range of communication standards and radar systems in low-altitude economic scenarios, thereby improving the versatility and feasibility of the method under different deployment conditions.

[0121] Figure 6 This is a block diagram of the overall structure of the multimodal joint spectrum allocation system for low-altitude economy provided by the present invention, as shown below. Figure 6 As shown, the system provided by the present invention consists of a signal acquisition module, a preprocessing module, a modal coding module, an intermediate fusion module, an information constraint module, a distribution consistency module, a situation map construction module, and a dynamic allocation module connected in sequence.

[0122] The signal acquisition module is deployed at ground base stations, towers, edge nodes, low-altitude monitoring stations, or airborne platforms to acquire multi-band communication signals and radar echo information. The preprocessing module performs synchronization, noise suppression, time-frequency transformation, time alignment, and spatial registration on the acquired basic observation data and outputs spatiotemporally aligned multimodal observation data.

[0123] The modality coding module includes coding branches constructed for communication standards and radar systems respectively, outputting communication modality latent features and radar modality latent features with a unified dimension; the intermediate fusion module performs at least one of the following processing on the two modality features at the latent feature layer: weighting, gating, attention or tensor fusion, to obtain fused features.

[0124] The information constraint module establishes a constraint relationship between the fused features and the target task output, improving the task relevance of the fused features and reducing redundant components; the distribution consistency module measures and minimizes the distribution difference between the potential features of the two modes, so that the two modes tend to align in a unified feature space.

[0125] The situation map construction module uses spatial location, time, and frequency as coordinate axes, mapping the fused features after distribution unification processing to a low-altitude spatial grid, and overlaying situation attribute information such as spectrum occupancy intensity, interference level, and link quality indicators to construct a three-dimensional situation map of low-altitude spectrum resources. The dynamic allocation module generates a dynamic spectrum resource allocation strategy based on this situation map and distributes it to execution objects such as low-altitude aircraft, ground base stations, or edge control platforms. Furthermore, the dynamic allocation module receives execution status feedback data from the execution objects and sends it back to the preprocessing module along with newly acquired multimodal observation data as input for the next cycle, enabling the system to form a complete closed loop in the data flow.

[0126] Figure 7 This is a flowchart illustrating the multimodal joint spectrum allocation method for low-altitude economy provided by the present invention, as shown below. Figure 7 As shown, the method provided by the present invention includes, in sequence, multimodal data acquisition, temporal and spatial alignment, feature encoding, intermediate fusion, information constraint optimization, distribution consistency processing, three-dimensional spectrum situation map generation, and dynamic spectrum allocation.

[0127] Specifically, the edge control platform first acquires multi-band communication signals and radar echo information from acquisition nodes distributed in the low-altitude airspace; then it performs timestamp calibration, spatial coordinate registration and noise suppression processing on the acquired multimodal observation data, and outputs spatiotemporally aligned multimodal observation data.

[0128] In the feature encoding stage, the communication signal and radar echo information are respectively sent to the corresponding mode encoding branches to obtain communication mode latent features and radar mode latent features with a unified dimension; then, in the intermediate fusion stage, the two mode features are fused in the latent space according to preset static weights or adaptive dynamic weights to obtain fused features.

[0129] The fused features are simultaneously processed in the information constraint optimization stage and the distribution consistency processing stage: the former improves the discrimination ability of the fused features for target tasks such as spectrum occupancy identification, interference detection, and link quality assessment by applying task relevance constraints and suppressing redundant components; the latter achieves mode alignment by minimizing the difference measurement results between the two mode distributions and outputs the fused features after distribution consistency.

[0130] The process then proceeds to the three-dimensional spectrum situation map generation stage. Using spatial location, time, and frequency as coordinate axes, the fused features after distribution consistency are mapped onto the low-altitude spatial grid and situation attribute information is superimposed to construct a three-dimensional situation map of low-altitude spectrum resources. After entering the dynamic spectrum allocation stage, a dynamic spectrum resource allocation strategy is generated based on this situation map and issued to the execution objects.

[0131] Furthermore, the method continuously acquires execution status feedback data returned by the execution object and newly acquired multimodal observation data in the new observation cycle, and uses both as input for the next cycle to return to the time-space alignment stage. This forms a closed-loop update link of perception, evaluation, decision-making and execution in the method process, enabling real-time tracking and dynamic adaptation to changes in the low-altitude electromagnetic environment and flight mission.

[0132] The multimodal joint spectrum allocation device provided by the present invention will be described below. The multimodal joint spectrum allocation device described below and the multimodal joint spectrum allocation method described above can be referred to in correspondence.

[0133] Figure 8 This is a schematic diagram of the multimodal joint spectrum allocation system provided by the present invention, as shown below. Figure 8 As shown, it includes: The modality coding module 810 is used to acquire multimodal observation data in the low-altitude airspace, and to perform feature coding on the multimodal observation data according to the mode to obtain communication mode potential features and radar mode potential features; wherein, the multimodal observation data includes multi-band communication signals and radar echo information; The intermediate fusion and constraint module 820 is used to perform intermediate fusion of the communication mode potential features and the radar mode potential features at the feature layer, and to apply information constraints to the intermediate fusion result to obtain fused features; wherein, the information constraints are used to enhance the correlation between the fused features and the target task output, and to reduce the redundant components in the fused features. The distribution unification module 830 is used to measure the distribution difference between the communication mode potential features and the radar mode potential features, and to perform distribution unification processing on the fused features based on the obtained distribution difference measurement results to obtain the distribution unification fused features; The situation map construction module 840 is used to construct a three-dimensional situation map of low-altitude spectrum resources based on the fusion features after the distribution is consistent, with spatial location, time and frequency as coordinate axes; The dynamic allocation module 850 is used to generate a dynamic spectrum resource allocation strategy based on the three-dimensional situation map of the low-altitude spectrum resources, and to send the dynamic spectrum resource allocation strategy to the execution object.

[0134] In this application, multi-modal observation data, including multi-band communication signals and radar echo information, is acquired in the low-altitude airspace. After feature encoding by mode, intermediate fusion is performed at the feature layer, establishing a coupling relationship between communication and sensing information in the latent space. Compared to merging results at the decision stage, this method retains more fine-grained complementary information related to spectrum occupancy, interference conditions, and link quality. By imposing information constraints on the intermediate fusion results, the fused features more centrally express the discriminative information related to the target task and suppress redundant components, improving the accuracy and robustness of sensing results in complex electromagnetic environments. Furthermore, by analyzing the latent features of communication and radar modes... The distribution differences between the modes are measured and the fused features are processed to achieve distribution consistency, which alleviates the representation offset problem caused by the differences in physical mechanisms and statistical laws of heterogeneous modes and enhances the model's generalization ability across devices, regions and environments. Furthermore, a three-dimensional situation map of low-altitude spectrum resources is constructed with spatial location, time and frequency as coordinate axes, and a dynamic spectrum resource allocation strategy is generated based on the three-dimensional situation map of low-altitude spectrum resources and then sent to the execution object. This enables the perception results to be directly converted into executable scheduling actions, forming a continuous processing link from perception to decision and then to execution, thereby improving the utilization efficiency of low-altitude airspace wireless resources, communication reliability and spectrum scheduling response capability.

[0135] Figure 9 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 9 As shown, the electronic device may include a processor 1010, a communications interface 1020, a memory 1030, and a communication bus 1040. The processor 1010, communications interface 1020, and memory 1030 communicate with each other via the communication bus 1040. The processor 1010 can call logical instructions in the memory 1030 to execute a multimodal joint spectrum allocation method. This method includes: acquiring multimodal observation data in the low-altitude airspace; performing feature encoding on the multimodal observation data according to mode to obtain communication mode potential features and radar mode potential features; wherein the multimodal observation data includes multi-band communication signals and radar echo information. The communication mode latent features and the radar mode latent features are fused at the feature layer, and information constraints are applied to the fusion result to obtain fused features; wherein, the information constraints are used to enhance the correlation between the fused features and the target task output, and to reduce the redundant components in the fused features. The distribution difference between the communication mode potential features and the radar mode potential features is measured, and the fused features are subjected to distribution unification processing based on the obtained distribution difference measurement results to obtain the distribution unified fused features; Based on the fusion features after the distribution is unified, a three-dimensional situation map of low-altitude spectrum resources is constructed with spatial location, time and frequency as coordinate axes. Based on the three-dimensional situation map of low-altitude spectrum resources, a dynamic spectrum resource allocation strategy is generated and the dynamic spectrum resource allocation strategy is sent to the execution objects.

[0136] Furthermore, the logical instructions in the aforementioned memory 1030 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0137] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the multimodal joint spectrum allocation method provided by the above methods. The method includes: acquiring multimodal observation data in the low-altitude airspace; performing feature encoding on the multimodal observation data according to mode to obtain communication mode potential features and radar mode potential features; wherein, the multimodal observation data includes multi-band communication signals and radar echo information. The communication mode latent features and the radar mode latent features are fused at the feature layer, and information constraints are applied to the fusion result to obtain fused features; wherein, the information constraints are used to enhance the correlation between the fused features and the target task output, and to reduce the redundant components in the fused features. The distribution difference between the communication mode potential features and the radar mode potential features is measured, and the fused features are subjected to distribution unification processing based on the obtained distribution difference measurement results to obtain the distribution unified fused features; Based on the fusion features after the distribution is unified, a three-dimensional situation map of low-altitude spectrum resources is constructed with spatial location, time and frequency as coordinate axes. Based on the three-dimensional situation map of low-altitude spectrum resources, a dynamic spectrum resource allocation strategy is generated and the dynamic spectrum resource allocation strategy is sent to the execution objects.

[0138] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the multimodal joint spectrum allocation method provided by the above methods. The method includes: acquiring multimodal observation data in low-altitude airspace; performing feature encoding on the multimodal observation data according to mode; and obtaining communication mode potential features and radar mode potential features; wherein the multimodal observation data includes multi-band communication signals and radar echo information. The communication mode latent features and the radar mode latent features are fused at the feature layer, and information constraints are applied to the fusion result to obtain fused features; wherein, the information constraints are used to enhance the correlation between the fused features and the target task output, and to reduce the redundant components in the fused features. The distribution difference between the communication mode potential features and the radar mode potential features is measured, and the fused features are subjected to distribution unification processing based on the obtained distribution difference measurement results to obtain the distribution unified fused features; Based on the fusion features after the distribution is unified, a three-dimensional situation map of low-altitude spectrum resources is constructed with spatial location, time and frequency as coordinate axes. Based on the three-dimensional situation map of low-altitude spectrum resources, a dynamic spectrum resource allocation strategy is generated and the dynamic spectrum resource allocation strategy is sent to the execution objects.

[0139] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0140] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0141] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multimodal joint spectrum allocation method, characterized in that, include: Multimodal observation data in the low-altitude airspace is acquired, and the multimodal observation data is feature-encoded according to mode to obtain communication mode potential features and radar mode potential features; wherein, the multimodal observation data includes multi-band communication signals and radar echo information; The communication mode latent features and the radar mode latent features are fused at the feature layer, and information constraints are applied to the fusion result to obtain fused features; wherein, the information constraints are used to enhance the correlation between the fused features and the target task output, and to reduce the redundant components in the fused features. The distribution difference between the communication mode potential features and the radar mode potential features is measured, and the fused features are subjected to distribution unification processing based on the obtained distribution difference measurement results to obtain the distribution unified fused features; Based on the fusion features after the distribution is unified, a three-dimensional situation map of low-altitude spectrum resources is constructed with spatial location, time and frequency as coordinate axes. Based on the three-dimensional situation map of low-altitude spectrum resources, a dynamic spectrum resource allocation strategy is generated and the dynamic spectrum resource allocation strategy is sent to the execution objects.

2. The multimodal joint spectrum allocation method according to claim 1, characterized in that, The distribution difference between the latent features of the communication mode and the latent features of the radar mode is measured, and the fused features are subjected to distribution unification processing based on the obtained distribution difference measurement results, including: The statistical distribution difference between the latent features of the communication mode and the latent features of the radar mode in a unified feature space is measured by at least one of the optimal transmission distance, kernel method distance and adversarial distribution matching method, and the distribution difference measurement result is obtained. With minimizing the distribution difference metric as the optimization objective, the representations of the communication mode latent features and the radar mode latent features in the unified feature space are aligned.

3. The multimodal joint spectrum allocation method according to claim 1 or 2, characterized in that, Measuring the distribution difference between the latent features of the communication modes and the latent features of the radar modes includes: The latent features of the communication modes and the latent features of the radar modes are respectively subjected to frequency domain transformation to obtain the transform domain features corresponding to each mode; The dominant frequency domain components in the transform domain features are retained, while the non-dominant frequency domain components are suppressed to obtain the frequency domain compressed features. A random mapping matrix with restricted isometry is used to perform dimensionality reduction projection on the frequency domain compressed features to obtain dimensionality reduction features located in a low-dimensional space. The distribution difference is approximated based on the dimensionality reduction features in the low-dimensional space to obtain the distribution difference measurement result.

4. The multimodal joint spectrum allocation method according to claim 1, characterized in that, The intermediate fusion of the communication mode latent features and the radar mode latent features at the feature layer includes: Based on the fusion weights corresponding to each mode, the latent features of the communication mode and the latent features of the radar mode are fused using at least one of the following methods: weighted fusion, gated fusion, attention fusion, and tensor fusion. The fusion weight is either a preset static weight or a dynamic weight adaptively generated based on dynamic weight generation factors. The dynamic weight generation factors include at least one of the following: current link quality, modal confidence, noise level, electromagnetic congestion in the flight area, and radar target tracking stability.

5. The multimodal joint spectrum allocation method according to claim 1, characterized in that, The target task output includes at least one of the following: spectrum occupancy identification result, interference detection result, available channel prediction result, low-altitude communication link quality assessment result, and flight area spectrum conflict early warning result; The reduction of redundant components in the fused features includes: suppressing duplicate information and local anomalous observation components between the communication mode latent features and the radar mode latent features.

6. The multimodal joint spectrum allocation method according to claim 1, characterized in that, After the step of issuing the spectrum resource dynamic allocation strategy to the execution object, the method further includes: The execution status feedback data generated by the execution object in response to the dynamic spectrum resource allocation strategy, as well as newly added multimodal observation data, are obtained; wherein, the execution status feedback data includes at least one of frequency band switching results, link quality changes, and interference avoidance results; The execution status feedback data and the newly added multimodal observation data are used together as the input data for the next cycle. The feature encoding, intermediate fusion, distribution unification processing, and construction of the three-dimensional situation map of low-altitude spectrum resources are performed in sequence to obtain the updated three-dimensional situation map of low-altitude spectrum resources. Based on the updated three-dimensional situation map of low-altitude spectrum resources, an updated dynamic spectrum resource allocation strategy is generated and distributed to the execution object.

7. The multimodal joint spectrum allocation method according to claim 1, characterized in that, Based on the fused features after distribution unification, a three-dimensional situation map of low-altitude spectrum resources is constructed using spatial location, time, and frequency as coordinate axes, including: The fused features after the distribution is made consistent are mapped onto a low-altitude spatial grid with spatial location, time and frequency as coordinate axes; Situational attribute information is superimposed on the low-altitude spatial grid to construct a three-dimensional situational map of the low-altitude spectrum resources; wherein, the situational attribute information includes at least one of spectrum occupancy intensity, interference level, link quality index, target activity density, and resource conflict degree.

8. The multimodal joint spectrum allocation method according to claim 1, characterized in that, The dynamic spectrum resource allocation strategy includes at least one of the following: priority channel recommendation, frequency band switching, link switching, transmit power adjustment, beam control, access method adjustment, interference avoidance path suggestion, and mission area resource reconfiguration. The target of execution includes at least one of low-altitude aircraft, ground base stations, and edge control platforms.

9. The multimodal joint spectrum allocation method according to claim 1, characterized in that, The step of performing feature encoding on the multimodal observation data according to each modality includes: For the multi-band communication signal, encoding is performed based on at least one of time-domain waveform, frequency-domain spectrum, time-frequency joint representation and modulation structure information to obtain the potential features of the communication mode. The radar echo information is encoded based on at least one of range dimension information, velocity dimension information, angle dimension information, micro-motion dimension information and trajectory sequence information to obtain the radar modal potential features; The feature encoding network used includes at least one or a combination of convolutional networks, temporal networks, graph structure modeling networks, and transform domain feature mapping networks.

10. The multimodal joint spectrum allocation method according to claim 1, characterized in that, Before performing feature encoding on the multimodal observation data according to each modality, the method further includes: The multimodal observation data is time-aligned and spatially registered to obtain spatiotemporally aligned multimodal observation data; The step of performing feature encoding on the multimodal observation data according to each mode includes: performing feature encoding on the spatiotemporally aligned multimodal observation data according to each mode to obtain the communication mode potential features and the radar mode potential features with a unified dimension.

11. The multimodal joint spectrum allocation method according to claim 1, characterized in that, The multi-band communication signals include at least one of the following: cellular mobile communication signals, WiFi signals, dedicated short-range communication signals, and UAV remote control and telemetry link signals; The radar echo information includes at least one of the following: target echo intensity, range, Doppler frequency shift, angular spectrum, micro-motion characteristics, and point trajectory characteristics.

12. A multimodal joint spectrum allocation system, characterized in that, include: The modal coding module is used to acquire multimodal observation data in the low-altitude airspace, and to perform feature coding on the multimodal observation data according to mode to obtain communication mode potential features and radar mode potential features; wherein, the multimodal observation data includes multi-band communication signals and radar echo information; The intermediate fusion and constraint module is used to perform intermediate fusion of the communication mode potential features and the radar mode potential features at the feature layer, and to apply information constraints to the intermediate fusion result to obtain fused features; wherein, the information constraints are used to enhance the correlation between the fused features and the target task output, and to reduce the redundant components in the fused features. The distribution unification module is used to measure the distribution difference between the communication mode potential features and the radar mode potential features, and to perform distribution unification processing on the fused features based on the obtained distribution difference measurement results to obtain the distribution unified fused features. The situation map construction module is used to construct a three-dimensional situation map of low-altitude spectrum resources based on the fused features after the distribution is consistent, with spatial location, time and frequency as coordinate axes; The dynamic allocation module is used to generate a dynamic spectrum resource allocation strategy based on the three-dimensional situation map of low-altitude spectrum resources, and to distribute the dynamic spectrum resource allocation strategy to the execution object.

13. An electronic device, characterized in that, It includes a memory and a processor; wherein the memory stores a computer program, and the processor, when executing the computer program, implements the steps of the multimodal joint spectrum allocation method as described in any one of claims 1 to 11.

14. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the multimodal joint spectrum allocation method as described in any one of claims 1 to 11.

15. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the multimodal joint spectrum allocation method as described in any one of claims 1 to 11.

Images